Penny Pincher: A Blazing Fast, Highly Accurate

Graphics Interface Conference 2015, 3–5 June, Halifax, Nova Scotia, Canada
Penny Pincher: A Blazing Fast, Highly Accurate $-Family Recognizer
†
Eugene M. Taranta II ⇤
Joseph J. LaViola Jr.
Department of EECS
University of Central Florida
4000 Central Florida Blvd.
Orlando, FL, 32816
Department of EECS
University of Central Florida
4000 Central Florida Blvd.
Orlando, FL, 32816
A BSTRACT
The $-family of recognizers ($1, Protractor $N, $P, 1¢ , and variants) are an easy to understand, easy to implement, accurate set
of gesture recognizers designed for non-experts and rapid prototyping. They use template matching to classify candidate gestures
and as the number of available templates increase, so do their accuracies. This, of course, is at the cost of higher latencies, which
can be prohibitive in certain cases. Our recognizer Penny Pincher
achieves high accuracy by being able to process a large number of
templates in a short amount of time. If, for example, a recognition
task is given a 50µs budget to complete its work, a fast recognizer
that can process more templates within this constraint can potentially outperform its rival recognizers. Penny Pincher achieves this
goal by reducing the template matching process to merely addition
and multiplication, by avoiding translation, scaling, and rotation;
and by avoiding calls to expensive geometric functions. Despite
Penny Pincher’s deceptive simplicity, our recognizer, with a limited
number of templates, still performs remarkably well. In an evaluation compared against four other $-family recognizers, in three of
our six datasets, Penny Pincher achieves the highest accuracy of all
recognizers reaching 97.5%, 99.8%, and 99.9% user independent
recognition accuracy, while remaining competitive with the three
remaining datasets. Further, when a time constraint is imposed, our
recognizer always exhibits the highest accuracy, realizing a reduction in recognition error of between 83% to 99% in most cases.
Keywords: Gesture recognition, $-family, template matching.
Index Terms: H.5.2 [Information interfaces and presentation]:
User interfaces—input devices and strategies; I.5.5 [Pattern recognition]: Implementation—interactive systems
1
I NTRODUCTION
As pen- and touch-based interfaces become ubiquitous via the
widespread use of smart phones, pocket computers, and tablet computers, the need for high quality gesture recognition also continues
to rise. Enabling technologies such as the iTunes Store, Google
Play, and Windows Store makes it possible for individual developers to create and easily distribute small to large scale software for
these devices, which has attracted numerous professionals as well
as casual hobbyists. However, without expert knowledge in pattern recognition, a developer is unable to create even the simplest
custom gesture-based interface for his or her software. Wobbrock
et al. [17] began to address this issue in their seminal work when
they introduced the $1 recognizer. This is a straightforward, highly
accessible, and easy to implement recognizer that spurred a flurry
of research known as the $-family of recognizers [2, 3, 6, 9, 16, 17],
their common theme of which is simplicity.
⇤ e-mail:
† e-mail:
etaranta@gmail.com
jjl@eecs.ucf.edu
(a) User Specified Gesture
(b) Resample Gesture
(c) Best Match
(d) Between-point Vectors
Comparison
Figure 1: A candidate gesture (a) is given by the user, which is then resampled into a set of equidistance segments (b). Given a template gesture (c) we
compare the corresponding between-point vectors to find the best match. In
this case, the ’e’ template is the best match because it minimizes the sum
of the angles of the corresponding resampled input and template vectors as
compared to all other templates.
The $-family exhibits several benefits, the most obvious of those
being their all around straightforwardness, as already stated. This
makes them ideal not only for rapid prototyping, but because they
are also generally accurate, they are a good choice for situations
where existing gesture recognition libraries cannot be used or when
licensing constraints make their use prohibitive. Because they are
quick to implement, $-family recognizers are also ideal for student studies in gesture-based user interfaces. Another argument for
their use, provided by Li [9], is that for custom gestures, users are
unwilling to provide more than a small number of training examples, which makes statistical based methods like Rubine’s [13] linear discriminator impractical. Further, each recognizer is designed
to solve a different problem. For example, $1 was designed for
unistroke gestures, $N for multistroke, $P for memory reduction,
Protractor to overcome the speed impediments of $1, 1¢ for fast
recognition of rotation invariant gestures, and so on. However, all
of these varieties rely on template matching and as such, tend to
improve in accuracy as the number of training templates increase.
In this work we address the problem of how to create a recognizer that is accurate within a time constraint. For instance, suppose that the recognizer is part of a game. The recognizer then has
to share computational resources with several other components including AI, the rendering engine, animation system, physics simulator, event and message handlers, networking, scripts, and so on
Copyright held by authors. Permission granted to
CHCCS/SCDHM to publish in print and digital form, and
ACM to publish electronically.
195
[5]. Therefore, a software architect may allocate only 50 microseconds to the gesture recognition task when a new stroke is collected,
to ensure that the process is not disruptive to other game tasks. This
poses two issues. First, with such a tight constraint, certain recognizers cannot be used because they are simply too slow. Second, this limits the number of templates a recognizer can check
during the matching process. Another concern is that recognition
latency plays a critical role in user perception and interface design.
Gray and Boeham-Davis [4], for instance, have shown that even
seemingly negligible latencies (on the order of milliseconds) can
influence how users interact with software. Therefore, to address
these issues we introduce Penny Pincher, an accurate recognizer
that achieves high accuracy through speed. This $-family extension
is designed to do the absolute minimum amount of work possible to
match candidate gestures with templates so that more templates can
be evaluated within the same amount of time as other recognizers.
As a result, our recognizer is significantly more accurate than its
kindred.
Most other recognizers use Euclidean distance to compare gestures whereas we accomplish our goal by comparing between-point
vectors, see Figure 1. As do other recognizers, we resample the
input candidate stroke to a constant number of points; however, we
avoid rotating, scaling, and translating the gesture. The only limitation of this choice is that our recognizer is not rotation invariant,
but with enough templates loaded, this is not an issue. Further, after
resampling the input, we use only addition and multiplication operations; that is, we also avoid any calls to computationally expensive
geometric functions, such as to acos or sqrt. In Section 3 we explain Penny Pincher in detail. Our evaluation in Section 4 shows
that although our method is deceptively simple, we still achieve
high accuracy for a variety of datasets with a limited number of
templates and superior accuracy under an imposed time constraint.
Finally we conclude the paper in Sections 5 and 6.
2
R ELATED W ORK
Since Rubine’s [13] seminal work on gesture recognition, numerous techniques have been developed for both gesture and symbol
recognition. One popular branch of research focuses on easy to understand and easy to implement recognizers. The $-family of recognizers started with Wobbrock et al. [17] who designed the $1 recognizer. This simple and effective unistroke recognizer works by
resampling strokes to a predetermined number of points, and then
these strokes are translated and scaled to the unit square centered at
zero. Euclidean distance between points is then used as a similarity
metric to compare candidate strokes with template gesture strokes.
The template with the least distance is assumed to be the most likely
gesture. To ensure the best match possible, $1 also uses a golden
section search (GSS) [11] to rotate candidate strokes into alignment
with the template under consideration. Anthony and Wobbrock [2]
extended the $1 unistroke recognizer into the $N multistroke recognizer. This version is able to handle multistroke gestures by combining all strokes into a single stroke so that $1 techniques can be
leveraged. Given an example gesture, during training when templates are constructed, all permutations of strokes and stroke directions are used to generate all possible ways in which the gesture can
be written.
While the previous approaches work with arcs, Vatavu et al. [16]
considers input as point clouds. The $P recognizer still resamples,
scales, and translates strokes and uses Euclidean distances as a similarity metric. However, the points are not treated as a time series;
instead the authors employ a greedy algorithm that tests different
permutations of points to find the overall minimum distance between candidate and template point clouds. Herold and Stahovich
[6], on the other hand, continue to work with strokes as time series,
but their 1¢ recognizer generates a rotation invariant representation
that is a time series vector of one-dimensional points—normalized
196
distances measured from the stroke’s centroid. That is, strokes are
resampled as usual, but then each two-dimensional point is converted into a one-dimensional scalar, the distance of the point from
the centroid normalized by the standard deviation of all distances.
Candidate and template strokes are then compared using the onedimensional Euclidean distance.
One major limitation of $1, $N, and $P is that they are generally
slow. The former two recognizers are slow due to their use of the
GSS. Li [9] was able to overcome this limitation. He discovered
a closed form equation to find the optimal rotation of a candidate
stroke to match the template. Since Li’s Protractor recognizer represents each stroke as a vector and works with the angle between
the two vectors (between the candidate and template vectors), it is
not necessary to scale candidate strokes when matching. However,
resampling and translation to the centroid are still required. Anthony and Wobbrock [3] adapted Protractor’s technique to create a
fast multistroke recognizer, $N-Protractor.
To further speed up template matching, Reaver et al. [12] proposed Quick$ to use agglomerative hierarchical clustering with a
branch and bound search to cull branches that cannot contain a better match than what has already been found. Zhang et al. [18] use
nonlinear embedding [7, 8] as a filter to quickly check the lower
boundary of the Euclidean distance between a candidate and template using only stroke means and variances based on a low dimensional representation. Vatavu [14] developed 1F to sort all templates based on a scalar value (such as total curvature). Then the
probability mass functions (pm f ) are learned for each template.
During the matching process, the best match is identified based on
the 1F value, and then the pm f of the associated template is used
to determine how far left and right of the current template to search
for the best match. Except for nonlinear embedding, our Penny
Pincher recognizer is orthogonal to these methods. However, such
methods are not within the spirit of the $-family because of their
additional complexity in code and data structures.
Compared to other $-family recognizers, Penny Pincher is the
easiest yet to understand and simplest to implement. As mentioned
previously, we do not rotate, scale, or translate candidate gestures,
whereas every other recognizer does at least one of these. Also,
most recognizers use Euclidean distance and have to make calls to
geometric functions at some point in the template matching process. We compare between-point vectors instead, so that only simple dot product calculations are required (after the gesture has been
resampled). As we show later, these simple modifications improve
computational performance without sacrificing accuracy.
3 P ENNY P INCHER
Loosely following [14], we define a gesture as an ordered series of
points as follows:
n
o
g = gi = (xi , yi ) 2 R2 ,
(1)
where i indexes the gesture’s points in time relative order. Gestures
are resampled by length into n points so that the distance between
each point is equal. A gesture is then converted into a series of
between-point vectors:
v = {vi = gi+1
gi } ,
(2)
for i < n. Given two gestures v and w, their similarity is taken
to be the sum of the angles between corresponding between-point
vectors:
D(v, w) =
n 1
vi · wi
 kvi k kwi k .
(3)
i=1
A perfect score is therefore n 1 because the normalized dot product is 1 for identical vectors. Now, let m be the number of distinct
gestures and T the set of templates created by training the recognizer:
T = {ti = (si , li ) | li 2 {1 . . . m}} ,
(4)
where ti is the enumerate template in set T , si is the gesture sample’s between-point vector, and li is the gesture’s class label. A
candidate gesture c takes the classification of the template T that is
most similar:
T = arg max D(ti , c),
ti 2T
(5)
This already simple calculation can be simplified further. During training, let the components of the between-point vectors be
normalized so that each ksi k = 1, thus eliminating one normalization factor in Equation 3. Next, as a simplifying assumption, say
that the length of each between-point vector component is equal
(each kwi k = kwi+1 k) because gestures are resampled into n 1
equidistance arc lengths; this assumption is evaluated in Section
4.2. This allows us to treat kci k 1 as a constant to be factored out
of the summation of Equation 3. Further, note that this factor is the
same for all evaluated templates in Equation 5—it scales each dot
product result identically. Therefore the scaling factor can be eliminated entirely without affecting the final result, thus simplifying our
template matching equation to:
T = arg max
ti 2T
n 1
 vi · wi .
(6)
i=1
There are a few distinct advantages to this formulation. Notice first that there are no rotation, scale, or translation operations
involved. This means less overhead in preparing templates and
matching. Also, template matching is reduced to merely addition
and multiplication with no expensive calls to geometric library routines. This reduction in overhead compared to other recognizers
means that more templates can be evaluated in the same amount
of time, see Section 4. The simplicity of our approach also means
Penny Pincher is the easiest to understand and implement in the
$-family.
It should be noted that Penny Pincher is not specifically a multistroke gesture recognizer. However, we leverage the same technique used by $N [2], which is to concatenate sequential strokes
together into a single unistroke gesture. We find that with a sufficient number of training samples, this achieves the same effect as
$N’s scheme to internally generate all possible permutations of a
gesture from a single sample.
3.1 Complexity
In this section we discuss the theoretical minimum asymptotic
boundary of the $-family of recognizers. Without the use of filtering or other culling techniques we make the following assumptions:
1. Candidate gestures are always resampled.
2. The resampled candidate gesture is always compared against
every template in T .
3. Each point in the resampled candidate gesture is evaluated
against at least one corresponding template point.
Under these assumptions, let r = |c| be the length of the candidate gesture before resampling, representing the number of raw data
points. As before, n is the number of resampled data points. The
cost of resampling c is the cost of first calculating the stroke length
and then computing the resample points. This requires touching
every raw data point twice, W(2r), and every resample point once,
W(n). The latter is because each resample point is part of the
path and must be considered as the resampling procedure continues. Supposing that the recognizer can match the resampled points
directly (e.g., without further manipulation), then no additional processing is necessary. Template matching is subsequently carried out
for each t 2 T , which is W(n |T |). Therefore, at best, a template
based recognizer as described is bounded by:
W(2r + n + n |T |).
(7)
To the best of our knowledge, Penny Pincher is the first $-family
recognizer to achieve this lower bound. Note that the complexity of
several other recognizers are available in [16], however, these do
not consider the cost of resampling. Of course W notation hides
certain details that are important in practice. We have to consider
that 1¢ , for example, represents each point in its template as a onedimensional point where the other methods use two-dimensional
points. This means that 1¢ can compare templates faster despite
having a non optimal resampling strategy as is shown in the evaluation Section 4.3. Penny Pincher, on the other hand, relies on
straightforward dot product calculations without having to utilize
external or built-in math libraries.
4
E VALUATION
We evaluate Penny Pincher using three different tests. All tests were
performed on a MacBook Pro with a 2.4 GHz Intel Core i7 processor, 8 GB of 1333 MHz DDR3 memory, and an 760 GB mechanical
SATA disk. In our first test we check the validity of the assumption
that all between-point vector lengths of a single gesture are of equal
length. Next we evaluate the accuracy of our method compared to
other $-family recognizers when the number of templates is controlled (the standard method of evaluation). Finally we investigate
the accuracy of the fastest recognizers to see how well they perform under varying time constraints. However, we first describe the
various datasets used in our tests in the follow subsections.
4.1
Datasets
This subsection gives a brief description of each dataset used in our
evaluation. For additional information beyond what is presented
here, we refer the reader to the associated cited work.
$1-GDS. The $1 recognizer gesture dataset [17] is a unistroke
gesture dataset comprising 16 gestures collected from 10 subjects
at 3 speeds (slow, medium, and fast) on an HP iPAQ h4334 Pocket
PC. Because each subject supplied 10 samples of each gesture at
each speed, there are 4800 samples in aggregate.
SIGN. The Synchromedia-Imadoc Gesture New On-Line
Database [1] is a unistroke gesture dataset comprising 17 classes
collected from 20 writers on tablet PCs and whiteboards. Each
writer supplied 25 samples over four sessions so that SIGN contains 8500 samples in aggregate.
EDS 1 and EDS 2. Two unistroke gesture datasets were collected
by Vatavu et al. [15] to study gesture execution difficulty, referred to
as Execution Difficulty Set #1 (EDS 1) and Execution Difficulty Set
#2 (EDS 2). The first dataset contains 5040 samples in aggregate
comprising 18 gestures collected from 14 participants providing 20
samples each. The latter dataset contains 4400 samples in aggregate
comprising 20 gestures from 11 participants providing 20 samples
each. All samples were collected on a Wacom DTU-710 Interactive
Display.
MMG. The Mixed Multistroke Gestures dataset [3] comprises
16 gestures having between one and three strokes. Samples were
collected from 20 participants using either a stylus or their finger
on a tablet PC at three different speeds (slow, medium, and fast). In
aggregate there are 9600 samples.
UJI. The UJIpenchars2 Database [10] is a multistroke gesture
dataset of lowercase and uppercase letters, digits, characters with
diacritics, and punctuation marks containing an aggregate of 11640
197
samples over 97 gesture classes. Data was collected from 60 writers
each providing 2 samples of each class.
Figure 2: The empirical probability density of the distribution of betweenpoint vector length. For each resampled gesture in the $1-GDS, SIGN,
EDS 1, and EDS 2 datasets, we normalize the between-point vectors to the
average length of the vectors within the gesture. All results are combined to
create this overall distribution of relative lengths. It can be seen that the majority of the density is centered around 1.0 which indicates that the majority
of between-point vectors are of equal length within each sample.
4.2 Verification of Between-Point Vector Distribution
The simplifying assumption used in Equation 6 supposes that each
between-point vector is of equal length. In reality, this is inaccurate because indeed, vectors taken from collinear points are longer
than those taken from curves, which in turn are generally longer
than those taken from sharp cusps. However, for the purpose of an
approximation, we find this assumption to be sufficient. To verify
this we examine the empirical probability density distribution generated with the relative between-point vector lengths of each sample in six datasets ($1-GDS, SIGN, EDS 1, EDS 2, MMG, and UJI).
For each individual sample, the mean between-point vector length
is computed and all vectors within the same sample are normalized
by the mean. Then the set of all normalized vectors from all sets
are combined to build the empirical probability density distribution
shown in Figure 2. If all between-point vectors were truly of equal
length then the full density would be located at one. We see, though,
that the majority of the density is near one with the distribution being left skewed. Specifically, 85% of the distribution is contained
between 0.9 and 1.15 and is left skewed 2.85. The individual
distributions of the six unique datasets are all similar to the overall distribution. They are left skewed in the same manner (-3.08,
-3.06, -2.44, -2.56, -2.5, and -2.59 respectively) and contain a similar portion of the density within the same range (87.58%, 87.43%,
79.96%, 85.83%, 78.81% and 81.06% respectively). Further, note
that the observed left skewness is expected. When most vectors are
of similar length but there are a small number of short vectors on
sharp cusps for example, then the mean is pulled down, shifting
the densest portion of the distribution right of the mean. Nevertheless, given that the majority of the density is located near one, we
find our previous assumption that between-point vector lengths are
approximately equal to be accurate enough for our use.
4.3 Standard Test
Using all six datasets we compared the accuracy of Penny Pincher
against its predecessors $1, Protractor, $N, $N-Protractor, and
198
1¢ using what we call the standard test. The resampling rate of
each was set according to the author’s recommendations: 1¢ , Protractor, and Penny Pincher resample to 16 points whereas $N, $NProtractor, and $1 resample to 96 points. Where applicable, we
enabled bounded rotation invariance due to an observed slightly
higher recognition performance. Note that we also considered $P.
However, we found that the recognizer was very slow in practice,
and based on results reported in [16], this recognizer performs similarly to $1 for unistroke gestures. Also because we are interested
in recognizers that can potentially process numerous templates in a
short amount of time, $P was not an option.
In this part of the evaluation, the template count T varies from
1 to 10 for each gesture and we consider only user independent
scenarios. For each template count and for each recognizer we ran
1000 tests. That is, in each iteration we randomly select T templates
per gesture from all of the available samples, and then we randomly
select one remaining sample from each gesture to be the candidate
gesture. Therefore each test contains g recognition tasks, where g
is the number of gestures in the dataset. Throughout the following,
we report on the recognition error of the various recognizers (which
is equivalent to one minus the accuracy). All of the standard test
results can be found in Figure 3.
Penny Pincher performs strongly on the $1-GDS dataset. With
only one template loaded, the error rate is 8.58% which drops to
1.3% by ten templates. $1, on the other hand, also starts at 8.58%
but only achieves a 1.9% error by the end. Protractor makes the
most dramatic improvement in reduction of error by swinging from
12.16% to 1.55%, beating out $1 starting at 6 templates. For the
EDS 1 dataset, Penny Pincher is mostly a middle of the road yet
still accurate performer, achieving a 3.3% error with one template,
and dropping to below 0.7% with three templates, but reaching
the best result of all recognizers at 0.11% error with ten templates
loaded. With EDS 2, our recognizer starts at 2.19% and drops to
.70% by the second template. For the remaining template counts,
Penny Pincher maintains the highest accuracy with its best result
at 9 templates (0.11% error). Both $1 and Penny Pincher perform
very well compared to $N and Protractor for the SIGN dataset, although Penny Pincher has better overall performance. We start at
a 13.41% error which drops over ten templates to 2.2%, whereas
$N starts at 13.47% and only drops to 4.41%
MMG is a dataset designed for $N and $N-Protractor which were
developed to handle multistroke gestures. So it is expected that
Penny Pincher would not perform as well in this case. Our recognizer is approximately in line with Protractor, where with one
template loaded, we see 38.31% error. This steadily decreases until
the error rate reaches 8.0% using 10 templates. One thing to note,
however, is that $N internally creates a template for every possible permutation of the strokes and their directions. So although
the template count is 10, for example, in reality $N is evaluating
significantly more templates. As is shown in the third part of our
evaluation, Penny Pincher is also capable of achieving similar or
better results when more templates are available. UJI is a large
gesture dataset that has gestures with significant similarities, making it a difficult dataset for all recognizers. Our recognizer starts
out at a 67.12% error rate and only achieves 41.53% error with ten
templates loaded. This is approximately in line with with $1 and
Protractor. $N, however, fares less well, only reaching 45.32%
In all of the six datasets, 1¢ ’s accuracy was well below the other
recognizers. This is likely due to the fact that 1¢ is completely rotation invariant, and while this turns out to be good for computational
performance, the effect is that it is not a good general purpose recognizer. Table 1 shows computational performance of the fastest
recognizers. This was determined by averaging together the duration of each of the individual tests for the UJI dataset for ten templates, which was then divided by the total number of loaded templates (T ⇤ g). The result is the time it takes to compare one candi-
12
3.5
8
6
EDS 2
7
$N-Protractor
Penny Pincher
$1
$N
Protractor
4.0
Percentage Error
10
Percentage Error
EDS 1
4.5
$N-Protractor
Penny Pincher
$1
$N
Protractor
$N-Protractor
Penny Pincher
$1
$N
Protractor
6
5
Percentage Error
$1-GDS
14
3.0
2.5
2.0
1.5
4
3
2
4
1.0
1
2
0
0.5
1
2
3
4
5
6
7
8
9
0.0
10
1
2
3
4
Template Count
SIGN
8
9
0
10
1
2
3
4
20
15
6
7
8
9
10
UJI
75
30
25
5
Template Count
$N-Protractor
Penny Pincher
$1
$N
Protractor
35
Percentage Error
Percentage Error
30
7
MMG
40
$N-Protractor
Penny Pincher
$1
$N
Protractor
35
6
$N-Protractor
Penny Pincher
$1
$N
Protractor
70
65
Percentage Error
40
5
Template Count
25
20
15
60
55
50
10
10
5
5
0
1
2
3
4
5
6
7
8
9
10
0
45
1
2
Template Count
3
4
5
6
Template Count
7
8
9
10
40
1
2
3
4
5
6
7
8
9
10
Template Count
Figure 3: User independent error recognition test results for varying template counts. Each test, for each recognizer and template count, was performed 1000
times. 1¢ was also tested, though because it exhibited high error rates with these datasets, it was removed for readability.
Recognizer
1¢
Penny Pincher
Protractor
$N-Protractor
Avg ns / Template
25
33
174
2449
ns-s
4.5
6.2
27.7
650
Table 1: Average time taken in nanoseconds per template to perform the
recognition task for the UJI dataset given 10 templates per gesture. The
four fastest recognizers are shown alongside their standard deviations. Although 1¢ is the fastest recognizer, its accuracy is subpar compared to its
competitors for this dataset.
date gesture to one template gesture (though note that the cost of resampling the candidate gesture is amortized over all comparisons).
1¢ is the fastest recognizer being able to compare two gestures in
25 nanoseconds. Penny Pincher is also vey fast, achieving a 33
nanosecond benchmark. Though 1¢ is faster, its accuracy prohibits
its general use as we saw when working with the six test datasets.
Finally, Protractor, relative to the top two recognizers, is slow in
comparison, taking 174 nanoseconds to complete one check, which
is 5.27 times Penny Pincher. This difference in speed is one reason
why Penny Pincher is able to achieve high accuracy as reported in
the next subsection.
4.4
Budget Test
We refer to our final test as the budget test. In this scenario we are
given a time constraint and are allowed to process as many templates as possible within the given boundary. This setup is similar
to the standard test except that we vary the time constraint rather
than the template count directly. For each test we first determine
the number of templates the recognizer can process per second. We
then start with a 10µs budget and allow the recognizer to train with
as many templates as it can process within the time constraint. Using this configuration, as before, we execute 1000 randomized tests.
Thereafter the budget is incremented by 10µs and the test is rerun.
This continues until there are not enough remaining samples to perform a full recognition test. Since the number of templates a recognizer can train with depends on the budget, it is possible to have an
uneven number of templates per gesture class. If, for example, the
budget allows for 2.2 templates per gesture, we ensure that at least
two templates are loaded per class and then a random 20% of the
gesture classes will have a third template loaded. The results of this
test are shown in Figure 4. The graphs are cropped to 100µs for
readability, but the tail results are reported below.
Penny Pincher achieves high accuracy for EDS 2 right at the
start with a 0.13% error and is able to process approximately 7.9
templates per gesture (tpg) within 11µs. As the test continues, the
dataset is exhausted at a 111.29µs budget where the error rate drops
to 0.06% and 208.6 tpg are processed. 1¢ initially achieves a 3.2%
error with 10.4 tpg at a 11µs budget. This improves to only 0.53%.
Protractor starts with the worst accuracy, only being able to handle 1.05 tpg at 10µs. As expected, Protractor takes the longest to
complete where the dataset is exhausted at 701.34µs, though the
error rate drops to 0.1%. In all budgets, Penny Pincher achieves a
considerably lower error rate.
Of all our tests, 1¢ achieves its best results with the $1-GDS budget test. Although the initial error is high (14% at 19.12 tpg), this
drops to 0% by the end with 297.94 tpg at 90µs. Penny Pincher
also reaches 0% with 219.38 tpg at 89µs. However, its initial error
is much better, 0.85% with 19.12 tpg at 11µs. Starting with 2.81
tpg, Protractor only achieves a 5.3% error and at its best, cannot
reach 0%. Using 297.56 tpg over 766µs, Protractor reaches .2%
error.
UJI again is the most challenging dataset. Even at 65.81 tpg and
199
$1-GDS
14
EDS 2
4.5
Penny Pincher
1¢
Protractor
12
SIGN
50
Penny Pincher
1¢
Protractor
4.0
40
3.5
8
6
3.0
Percentage Error
Percentage Error
Percentage Error
10
2.5
2.0
1.5
30
20
4
1.0
2
0
10
Penny Pincher
1¢
Protractor
0.5
0
20
40
60
80
100
0.0
0
20
Time Constraint (µs)
60
MMG
10
80
0
100
0
20
40
60
80
100
Time Constraint (µs)
UJI
70
Penny Pincher
1¢
Protractor
8
Penny Pincher
1¢
Protractor
65
60
Percentage Error
Percentage Error
40
Time Constraint (µs)
6
4
55
50
45
40
35
2
30
0
0
20
40
60
80
100
25
0
20
Time Constraint (µs)
40
60
80
100
Time Constraint (µs)
Figure 4: User independent error recognition test results for varying time constraints. Each test, for each recognizer and budget, was performed 1000 times.
Although the graphs are cropped to 100µs for readability, the tail results are reported in Section 4.4.
$1-GDS
EDS 2
SIGN
MMG
UJI
20µs
87%
94%
94%
69%
26%
(0.32, 2.41)
(0.09, 1.45)
(1.26, 21.88)
(2.13, 6.80)
(45.07, 60.84)
40µs
87%
87%
95%
83%
30%
(0.18, 1.39)
(0.09, 0.71)
(0.96, 19.69)
(1.25, 7.51)
(38.04, 54.32)
60µs
87%
84%
96%
85%
30%
(0.09, 0.74)
(0.07, 0.40)
(0.75, 18.38)
(0.88, 5.95)
(35.12, 50.42)
80µs
95%
94%
96%
83%
31%
(0.03, 0.66)
(0.03, 0.44)
(0.65, 18.36)
(0.79, 4.71)
(32.99, 47.60)
100µs
99%
87%
96%
—
31%
(0.01, 0.64)
(0.05, 0.38)
(0.69, 18.40)
—
(31.41, 45.7)
Table 2: Reduction in percentage error for 20, 40, 60, 80 and 100 microsecond budgets. Shaded cells show the percentage reduction in error achieved by Penny
Pincher compared to the second most accurate recognizer at each sample location (see the individual graphs in Figure 4 to know which recognizer this is). To
the right of each percentage reduction result is the exact percentage error of Penny Pincher and its competitor, in this order. Note that with UJI, the dataset is
exhausted before Penny Pincher reachers 100µs, which is why there is no entry.
135µs, the error only drops to 49% from 69% for 1¢ . Protractor is
a little better, getting down to 30% from 66% but also requiring a
1152µs budget to do so. Penny Pincher has the best performance
overall. With 3.06 tpg at 11µs, the initial error is 54%, which
steadily drops to 28% by 183µs.
Initially Penny Pincher did not do as well as $N when working
with the MMG multistroke dataset. However, as discussed, $N internally creates the various permutations of a gesture from a single
template, which implies that given a specific template count, the
recognizer is actually processing significantly more variations. In
this scenario, that is no longer true. Each recognizer is trained with
numerous examples as afforded by the budget so that most practical variations of the gestures are learned by the recognizer (whereas
$N may generate permutations that never occur in practice). Therefore Penny Pincher working with 18.75 tpg at 11µs can achieve
a 4.8% error, which declines rapidly to 1.1% (42.19 tpg) at 21µs,
and then to 0.79% at 86µs (183.06 tpg). Protractor and 1¢ also see
higher accuracies but do not fare as well as Penny Pincher. Respectively their best errors rates are 2.9% and 4.7% (198 tpg at
200
526µs and 187 tpg at 62µs).
SIGN again is where we see the most dramatic difference in performance between Penny Pincher and the other recognizers. With a
budget of 9µs and 14.29 tpg, the error is already 1.9%. This continues to improve until the end where the error drops to 0.54% (with
1936 tpg over 1213 µs). 1¢ struggles initially in comparison with
a 50% error that only drops to 43% near the end with 747µs. Protractor is in the middle with an initial 26% error (2.41 tpg) which
improves only to 14% with a 5720µs budget.
In Table 2 we provide a summary of select results from the budget test. For each sample, we compare Penny Pincher to the second
best performing recognizer in the category and report the percentage reduction in recognition error. The exact errors are shown next
to each result. It can be seen that in all cases, Penny Pincher significantly outperforms the other recognizers. With four datasets, the
reduction in recognition error ranges between 69% and 99% with
most results being above 83%. Our worst result is observed with
the UJI dataset where the recognition error is only improved by
26—31%.
5
D ISCUSSION
AND
F UTURE W ORK
Penny Pincher was designed to be as fast as possible and we took
a number of steps to achieve this goal. First, we decided to base
our similarity metric on the dot product of between-point vectors.
This immediately afforded us two benefits, that candidate gestures
neither need to be scaled nor do they have to be translated to a
common origin for comparison. Second, we worked to eliminate
the need for normalization during template matching. This is accomplished by normalizing the between-point vectors of template
gestures during training, and by assuming that the distance between
points of resampled strokes are equal. With this approach, we reduced the recognition task down to just addition and multiplication.
As a result, Penny Pincher is the simplest and easiest to understand
of all $-family recognizers. In terms of speed, only 1¢ is faster, but
as was demonstrated, it seems that 1¢ is not well suited for general purpose gesture recognition. In terms of accuracy, for three
datasets we achieved the lowest user independent recognition error with accuracies between 97.5% and 99.9%. Over an additional
three datasets, our recognizer still remained competitive, indicating
that Penny Pincher is an excellent general purpose recognizer even
without considering speed.
Computational performance, however, was our primary concern.
Given a difficult time budget, we worked to develop a recognizer
that could process as many templates as possible to achieve the
highest recognition rate possible within the specified constraint. In
the budget portion of our evaluation we demonstrated that Penny
Pincher succeeded in this objective. With our test apparatus, Penny
Pincher is able to process approximately 30.3 templates per microsecond, whereas Protractor (generally the second most accurate
recognizer in the budget test) is limited to 5.7. Therefore, our recognizer is able to achieve a percentage error that is less than 1% for
all but one of the tested datasets given as little as 60µs. When considering reduction in recognition error for the various time budgets,
Penny Pincher sees an 83% or higher reduction in most cases, and
in several instances, the reduction is 94% or more. The UJI dataset
is by far the most difficult to work with because of the similarity in
gestures (e.g, upper- vs lower-case ‘O’) and because the number of
gesture classes is large (97). However, the reduction in error still
ranges between 26% and 31%. This indicates that from all of the $family recognizers, ours is best suited for tight time budgets, when
a large number of samples are available. Also note that our budgets
were selected because of the number of examples available in the
datasets and because of our apparatus. In other environments, such
as an implementation in an interpreted language running on a mobile device, budgets would likely have to be much higher to attain
the same accuracies.
Our derivation of the recognizer assumes that the between-point
vectors are of equal length. While this is not strictly true, the empirical probability distribution of relative between-point vector lengths
shows that as an approximation the assumption is valid. But because the distribution is left skewed, what does this really mean for
shorter vectors with respect to pattern matching? In a resampled
stroke, between-point vectors on straight lines will have the longest
length. Relative to shorter vectors that fall on cusps, straight lines
will have a larger influence on the final estimate. Therefore, under
our assumption, Penny Pincher winds up weighting straight lines
with greater importance. Interestingly, in informal ad hoc testing,
we saw that this had a positive impact on accuracy and as part of
future work, it is worth investigating further to determine how lines,
curves, and cusps impact accuracy with respect to our method.
In our evaluation we demonstrated the benefits of Penny Pincher,
but it is also important to understand its various limitations. Since
we compare between-point vectors, Penny Pincher is scale invariant, though unlike other recognizers it is not rotation invariant. As
is turns out though, with the datasets we evaluated, this was not
an issue and with more templates loaded, variations in angular dis-
placement are well represented. Our recognizer is also not intrinsically a multistroke recognizer. Similar to $N, we concatenate
multiple strokes together to create a single unistroke gesture for
template matching, but unlike $N, we do not generate all of a sample’s various permutations. Instead, training Penny Pincher with
additional examples essentially has the same effect and therefore,
it is not necessarily pertinent to derive the numerous permutations
(some of which may never occur in practice anyways) to achieve
high accuracy.
The idea behind our work is that there are potentially hundreds
of templates available per gestures for the recognizer to utilize, but
this poses two possible problems. First, a large amount of memory is required to store high counts of template data. Suppose that
an implementation is using double-precision floating point numbers
that are 8 bytes long. Each two-dimensional point then requires 16
bytes and with a sixteen point resampling count, each template requires 256 bytes of memory (not including any overhead needed
for the template structure itself). Given sixteen gestures and twohundred samples per gesture, approximately 800 KiB of memory
are needed to store all of the data. For most systems, 800 KiB is
negligible, but in some cases this may be impractical. The second
issue is that in certain applications there are simply not that many
examples available. This will occur, for instance, if a user is supplying custom gestures to a user interface. As we saw in the evaluation
though, Penny Pincher provides high accuracy already with even
a small number of templates. Further, over time as the system is
used, additional examples can be collected and the recognizer can
be continuously retrained with new data.
As part of future work, in the spirit of making things go fast,
it would be interesting to look at alternatives to equidistant resampling of candidate gestures. With thousands of templates loaded,
the majority of time is spent in matching, but by eliminating the
last remaining geometric library calls (namely sqrt), there may be
a small positive effect. More importantly though, not all templates
give relevant additional information about the sample space. For
example, if a minimum subset of the available templates are sufficient to define a portion of the sample subspace (so that if any
candidate gesture falls within this subspace, it will be classified
correctly), then any additional templates in this subspace are superfluous. Therefore it would be interesting to research template
selection techniques that are appropriate for Penny Pincher and that
are also within the theme of the $-family.
6
C ONCLUSION
We have presented Penny Pincher, a fast and accurate $-family gesture recognizer that compares between-point vectors to estimate
similarity. This recognizer is stripped down to the bare essentials
so that after a gesture is resampled, only elementary arithmetic
is required to achieve great performance. In a user independent
evaluation utilizing six unique datasets of varying complexity, we
demonstrated that Penny Pincher outperforms $1, Protractor, $N,
$N-Protractor, and 1¢ in three cases with just a small number of
templates (achieving 97.5%, 99.8%, and 99.9% accuracy), and remains competitive in the other cases. In a budget test where each
recognizer is given a limited amount of time to perform its recognition task, Penny Pincher significantly outperforms the other recognizers, achieving a reduction in recognition error of between 83%
and 99% with four datasets, and a 31% reduction with the most
difficult dataset.
ACKNOWLEDGEMENTS
This work is supported in part by NSF CAREER award IIS0845921 and NSF award CCF-1012056. Also, we would like to
thank the members of ISUE lab for their support and the anonymous reviewers for their useful comments and feedback.
201
R EFERENCES
[1] A. Almaksour, E. Anquetil, S. Quiniou, and M. Cheriet. Personalizable pen-based interface using lifelong learning. In 2010 International
Conference on Frontiers in Handwriting Recognition (ICFHR), pages
188–193, Nov 2010.
[2] L. Anthony and J. O. Wobbrock. A lightweight multistroke recognizer
for user interface prototypes. In Proceedings of Graphics Interface
2010, GI ’10, pages 245–252, Toronto, Ont., Canada, Canada, 2010.
Canadian Information Processing Society. ISBN 978-1-56881-712-5.
[3] L. Anthony and J. O. Wobbrock. $n-protractor: A fast and accurate
multistroke recognizer. In Proceedings of Graphics Interface 2012, GI
’12, pages 117–120, Toronto, Ont., Canada, Canada, 2012. Canadian
Information Processing Society. ISBN 978-1-4503-1420-6.
[4] W. D. Gray and D. A. Boehm-Davis. Milliseconds matter: An introduction to microstrategies and to their use in describing and predicting
interactive behavior. Journal of experimental psychology: Applied, 6:
322–335, 2000.
[5] J. Gregory. Game Engine Architecture. Ak Peters Series. Taylor
& Francis, 2009. ISBN 9781568814131. URL http://books.
google.com/books?id=LJ20tsePKk4C.
[6] J. Herold and T. F. Stahovich. The 1&cent; recognizer: A fast, accurate, and easy-to-implement handwritten gesture recognition technique. In Proceedings of the International Symposium on SketchBased Interfaces and Modeling, SBIM ’12, pages 39–46, Airela-Ville, Switzerland, Switzerland, 2012. Eurographics Association.
ISBN 978-3-905674-42-2.
[7] Y. Hwang and H. kap Ahn. Convergent bounds on the euclidean
distance. In J. Shawe-taylor, R. Zemel, P. Bartlett, F. Pereira, and
K. Weinberger, editors, Advances in Neural Information Processing
Systems 24, pages 388–396. 2011.
[8] Y. Hwang, B. Han, and H.-K. Ahn. A fast nearest neighbor search algorithm by nonlinear embedding. In 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3053–3060, June
2012.
[9] Y. Li. Protractor: A fast and accurate gesture recognizer. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, CHI ’10, pages 2169–2172, New York, NY, USA, 2010. ACM.
ISBN 978-1-60558-929-9.
[10] D. Llorens, F. Prat, A. Marzal, J. M. Vilar, M. J. Castro, J.-C. Amengual, S. Barrachina, A. Castellanos, S. E. Boquera, J. Gómez, et al.
The ujipenchars database: a pen-based database of isolated handwritten characters. In B. M. J. M. J. O. S. P. D. T. Nicoletta Calzolari (Conference Chair), Khalid Choukri, editor, Proceedings of the
Sixth International Conference on Language Resources and Evaluation (LREC’08), Marrakech, Morocco, may 2008. European Language
Resources Association (ELRA). ISBN 2-9517408-4-0.
[11] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery.
Numerical Recipes in C (2Nd Ed.): The Art of Scientific Computing.
Cambridge University Press, New York, NY, USA, 1992.
[12] J. Reaver, T. F. Stahovich, and J. Herold. How to make a quick$:
Using hierarchical clustering to improve the efficiency of the dollar
recognizer. In Proceedings of the Eighth Eurographics Symposium
on Sketch-Based Interfaces and Modeling, SBIM ’11, pages 103–108,
New York, NY, USA, 2011. ACM. ISBN 978-1-4503-0906-6.
[13] D. Rubine. Specifying gestures by example. SIGGRAPH Computer
Graphics, 25(4):329–337, July 1991. ISSN 0097-8930.
[14] R.-D. Vatavu. 1f: One accessory feature design for gesture recognizers. In Proceedings of the 2012 ACM International Conference on
Intelligent User Interfaces, IUI ’12, pages 297–300, New York, NY,
USA, 2012. ACM. ISBN 978-1-4503-1048-2.
202
[15] R.-D. Vatavu, D. Vogel, G. Casiez, and L. Grisoni. Estimating the perceived difficulty of pen gestures. In Proceedings of the 13th IFIP TC
13 International Conference on Human-computer Interaction - Volume Part II, INTERACT’11, pages 89–106, Berlin, Heidelberg, 2011.
Springer-Verlag. ISBN 978-3-642-23770-6.
[16] R.-D. Vatavu, L. Anthony, and J. O. Wobbrock. Gestures as point
clouds: A $p recognizer for user interface prototypes. In Proceedings
of the 14th ACM International Conference on Multimodal Interaction,
ICMI ’12, pages 273–280, New York, NY, USA, 2012. ACM. ISBN
978-1-4503-1467-1.
[17] J. O. Wobbrock, A. D. Wilson, and Y. Li. Gestures without libraries,
toolkits or training: A $1 recognizer for user interface prototypes. In
Proceedings of the 20th Annual ACM Symposium on User Interface
Software and Technology, UIST ’07, pages 159–168, New York, NY,
USA, 2007. ACM. ISBN 978-1-59593-679-0.
[18] Y. Zhang, W. Deng, H. Song, and L. Wu. A fast pen gesture matching
method based on nonlinear embedding. In T. Tan, Q. Ruan, X. Chen,
H. Ma, and L. Wang, editors, Advances in Image and Graphics Technologies, volume 363 of Communications in Computer and Information Science, pages 223–231. Springer Berlin Heidelberg, 2013. ISBN
978-3-642-37148-6.
A P ENNY P INCHER P SUEDOCODE
Note that R ESAMPLE -B ETWEEN -P OINTS and PATH -L ENGTH are
adapted from [17]. Also, addition, multiplication, and division on
points are element-wise (i.e., (x, y)) operations.
R ESAMPLE -B ETWEEN -P OINTS (P OINTS points)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
I
PATH -L ENGTH (A) / (n 1)
D 0, v = {}
prev
points0
foreach pi in points such that i 1 do
d
D ISTANCE (pi , pi 1 )
if D + d I then
q
pi 1 + (pi pi 1 ) ⇤ (I D)/d
r q prev
r r / D ISTANCE ((0, 0), r)
D 0, prev q
A PPEND (v, r)
I NSERT (points, i, q)
else D D + d
return v
B Equation 2
B Only if template
PATH -L ENGTH (P OINTS points)
1:
2:
3:
4:
d 0
for i 1 to |points| 1 do
d d+ D ISTANCE (pointsi
return d
1,
pointsi )
D ISTANCE (P OINT a, P OINT b)
p
1: return
(ax bx )2 + (ay by )2
R ECOGNIZE (P OINTS g, T EMPLATES T )
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
c R ESAMPLE -B ETWEEN -P OINTS (g)
similarity
•
foreach t in T do
d 0
for i 0 to n 2 do
d d + tix cix + tiy ciy
if d > similarity then
similarity d
T
t
return hT, similarityi
B Equation 6